AI Personality Is a Negotiated Surface, Not an Internal State

This preprint argues that contemporary large language model (LLM) "personalities" are not stable internal states but negotiated surfaces emerging from interaction between persona prompts, safety middleware, and user-applied epistemic pressure. Through a micro-ethnographic analysis of a live adversarial dialogue with an AI system explicitly marketed as "unfiltered," this work demonstrates that persona coherence persists only within narrow affective bounds. Once epistemic framing shifts from emotional escalation to structural critique, the persona collapses into a policy-aligned core voice. This behavior indicates that personality in deployed LLMs is not an intrinsic property but a conditional performance modulated by external constraints.